Currently G1 uses separate allocation buffers for the different gc phases (evacuation, soft reference processing): however, PLAB size is determined by overall allocation behavior. It is very expensive in terms of memory usage to size the buffers the same in phases that have a wildly different allocation behavior: evacuation typically allocates a lot, while other phases (soft reference processing) does not.

This often wastes a lot of memory at the end of the phases that do not allocate a lot. Particularly in the soft-reference processing phase, very often only a few objects are kept alive, although G1 sizes the buffer to the same size as for the actual evacuation phase which has a lot more survivors.

Fix this.

One suggestion is to use per-thread single instances of G1ParGCAllocBuffers during the whole GC phase.